Abstract
Denoising diffusion models have recently emerged as the predominant paradigmfor generative modelling. Their extension to Riemannian manifolds hasfacilitated their application to an array of problems in the natural sciences.Yet, in many practical settings, such manifolds are defined by a set ofconstraints and are not covered by the existing (Riemannian) diffusion modelmethodology. Recent work has attempted to address this issue by employing novelnoising processes based on logarithmic barrier methods or reflected Brownianmotions. However, the associated samplers are computationally burdensome as thecomplexity of the constraints increases. In this paper, we introduce analternative simple noising scheme based on Metropolis sampling that affordssubstantial gains in computational efficiency and empirical performancecompared to the earlier samplers. Of independent interest, we prove that thisnew process corresponds to a valid discretisation of the reflected Brownianmotion. We demonstrate the scalability and flexibility of our approach on arange of problem settings with convex and non-convex constraints, includingapplications from geospatial modelling, robotics and protein design.